Implementation of Publish Subscribe Message Exchange interface #1676

Signed-off-by: Dennis Potman <[email protected]> Co-authored-by: Martijn Reicher <[email protected]> Co-authored-by: Erik Boasson <[email protected]> Co-authored-by: Thijs Miedema <[email protected]>

Signed-off-by: Dennis Potman <[email protected]>

…igin member in struct loaned_sample to enum with kind and psmx-instance ptr Signed-off-by: Dennis Potman <[email protected]>

Signed-off-by: Dennis Potman <[email protected]>

…not enabled Signed-off-by: Dennis Potman <[email protected]>

Signed-off-by: Dennis Potman <[email protected]>

…stances Signed-off-by: Dennis Potman <[email protected]>

…as a key Signed-off-by: Dennis Potman <[email protected]>

Signed-off-by: Dennis Potman <[email protected]>

…lable based on PSMX and add op for getting supported-feature to PSMX instance ops Signed-off-by: Dennis Potman <[email protected]>

…ferent headers Signed-off-by: Dennis Potman <[email protected]>

Signed-off-by: Dennis Potman <[email protected]>

The cdr_identifier field in psmx meta-data should have a valid encapsulation identifier, not just the XCDR version (or undefined in case of a raw sample). This commit introduces a vendor specific RTPS encapsulation kind for a raw sample in native endianess, and it fixes the use of the cdr_identifier field. Signed-off-by: Dennis Potman <[email protected]>

Replace the writes in psmx test with the with_timestamp variant, and a check on the source timestamp is added. The cdds psmx implementation did not set the timestamp when receiving a sample, so that is also fixed. Signed-off-by: Dennis Potman <[email protected]>

If the plugin doesn't support keys then we have to skip tests with topic types that have keys. Signed-off-by: Martijn Reicher <[email protected]>

Due to the limitations of strtoull the string retrieved from iox needs to be chopped up into 2 character chunks, then converted Signed-off-by: Martijn Reicher <[email protected]>

Since the hash is something that is not transferable between different CycloneDDS processes, copying this is wrong, and recalculation is necessary Signed-off-by: Martijn Reicher <[email protected]>

Not acting on or storing return values is dangerous in the best of cases, but in this, the functions called may invoke code not inside CycloneDDS, and the predictability of their results is therefore also not constant In this case, non-OK return codes will therefore now be an early exit condition and a return of the function Signed-off-by: Martijn Reicher <[email protected]>

The functions dds_is_loan_available and dds_loan_sample are now deprecated in the documentation, but this was not yet implemented for their attributes. This caused CLang issues as warnings are treated as errors. Signed-off-by: Martijn Reicher <[email protected]>

Signed-off-by: Erik Boasson <[email protected]>

For refcounted data, calling unref rather than simply freeing the data is the correct way. The only (real) call to dds_loaned_sample_free was in the error path of requesting a writer loan, so no reason to keep it around. Signed-off-by: Erik Boasson <[email protected]>

All dead code: the back refs are not used anywhere and the loan pools are only manipulated while holding the lock of the reader/writer that owns the pool, so the mutex is not needed. Signed-off-by: Erik Boasson <[email protected]>

Signed-off-by: Erik Boasson <[email protected]>

Not just those in domain 0. Signed-off-by: Erik Boasson <[email protected]>

Signed-off-by: Erik Boasson <[email protected]>

They always come in pairs, and knowing whether the topic is keyed or not is necessary if the underlying service does not natively supports keys but does support blocking: in that case, keyed volatile data is ok, but non-volatile data is not. Signed-off-by: Erik Boasson <[email protected]>

Signed-off-by: Erik Boasson <[email protected]>

They induce a lot of load because of the huge amount of discovery, and I suspect that causes them to fail on CI. Especially the Cyclone-based PSMX implementation is nasty. Signed-off-by: Erik Boasson <[email protected]>

Signed-off-by: Erik Boasson <[email protected]>

This (inefficiently) allows for multiple PSMX instances in the checking code, but there are many more places where multiple PSMX instances need extra work. This makes the "local_psmx" flag on proxy endpoints unused. Arguably, ddsi_plist do_locator should filter out irrelevant PSMX locators, why carry them round if you're never going to use them? Signed-off-by: Erik Boasson <[email protected]>

Signed-off-by: Erik Boasson <[email protected]>

Pointers to shared copies of data (shared across data readers, possibly across processes if using PSMX) should only be returned by the "with loan" variants. This commit takes care of that and documents the behaviour in much greater detail, especially also with respect to the use of the pointer array that read/take use for returning samples. Signed-off-by: Erik Boasson <[email protected]>

Signed-off-by: Erik Boasson <[email protected]>

Read (non-_wl) must always result in private copies (as per the documentation), so that is easy to check. Read-with-loan may or may not return shared copies, and which of the two happens is highly implementation dependent, both of Cyclone's core and of a PSMX plugin. This test checks it against current behaviour, in particular verifying zero-copy and shared loans of self-contained data types in combination with Iceoryx. Signed-off-by: Erik Boasson <[email protected]>

With this commit, dds_read/dds_take and all variants now return pointers to shared copies of data if this is possible (which it not always is). As a consequence, modifying the contents of the loans will have possibly disastrous effects, whereas before one could away with it. If a private copy is required, the application must now allocate the memory itself (if there are any embedded pointers, those must be null pointers). One must now do: void *sampleptr = dds_alloc (sizeof (T)); dds_take (reader, &sampleptr, sampleinfo, 1, 1); ... dds_sample_free (sampleptr, &T_desc, DDS_FREE_ALL); to get a private copy in heap-allocated memory. (Obviously, other allocation/freeing routines are ok as well.) This is a breaking change and it has not been done lightly. The considerations are: - This really only breaks applications that modify "loaned" memory, which has always been dodgy behaviour at best, and is assumed to be extremely rare in application code. - The majority of applications will now benefit from zero-copy behaviour where it is supported, and without having to worry about which API function is the appropriate one to use. - It reduces the API surface. Signed-off-by: Erik Boasson <[email protected]>

Signed-off-by: Erik Boasson <[email protected]>

With the concomitant changes to the circumstances under which an entry is set to a null pointer and corresponding updates to the documentation. Signed-off-by: Erik Boasson <[email protected]>

This makes the behaviour consistent between readers and writers. Signed-off-by: Erik Boasson <[email protected]>

Signed-off-by: Erik Boasson <[email protected]>

Note that this requires the first/next reader operations passed in to check the flag and skip readers that have it set, but that is perfectly reasonable. Signed-off-by: Erik Boasson <[email protected]>

Signed-off-by: Dennis Potman <[email protected]>

with newline (to get proper output on CI for failing Windows builds) Signed-off-by: Dennis Potman <[email protected]>

When we switch from a dds to a psmx writer, we start with just one local reader and expect that there are no locators in the writer's address set, but: - if there is no unacknowledged data in the writer at the time it gets deleted, and - the discovery of the demise of the old readers may not have taken place yet at the time of creation of the writer, and - consequently the writer was matched with the old readers, and then - the discovery of said demise happens just prior to fetching the list of matched subscriptions, and - also prior to the GC performing the actual unmatching of the proxy readers and the waitset in "allmatched" is spinning/happens to wake up at the right time, then the test code will see the correct set of readers, but there are still some ports lingering around. It is very unlikely, but it can happen. Signed-off-by: Erik Boasson <[email protected]>

* Track current value for a few seconds before freezing it as the reference value * Add a custom allocator wrapper that tracks exact memory usage, because RSS turns out to be too fragile (RSS is also still available) * Tweak sanity check script to use a tight live memory bound and a generous RSS bound Signed-off-by: Erik Boasson <[email protected]>

Signed-off-by: Dennis Potman <[email protected]>

Signed-off-by: Erik Boasson <[email protected]>

* Add ddsi_sertype_init_props * Clean up sertype flags * Rename IS_FIXED_SIZE to IS_MEMCPY_SAFE * Delete useless "psmx endpoint serialization required" Signed-off-by: Erik Boasson <[email protected]>

Signed-off-by: Erik Boasson <[email protected]>

Integrating (for example) application code using Iceoryx with Cyclone DDS + PSMX + Iceoryx requires that the application knows the type name. So we must pass it to the plug-in. It seems to be best to use the actual type name by default for this reason. Signed-off-by: Erik Boasson <[email protected]>

Signed-off-by: Dennis Potman <[email protected]>

* ENABLE_SHM => ENABLE_ICEORYX * No code referencing Iceoryx remaining in the core, except in code for supporting backwards compatibility * No more references to iceoryx_binding_c * CI builds use "iceoryx" instead of "psmx_iox" Signed-off-by: Erik Boasson <[email protected]>

For keyless topics, a serialized key has length 0, for which dds_take leaves the _buffer field a null pointer. Signed-off-by: Erik Boasson <[email protected]>

Signed-off-by: Erik Boasson <[email protected]>

Signed-off-by: Dennis Potman <[email protected]>

…ation Signed-off-by: Dennis Potman <[email protected]>

Update (and rename) dds_loan_shared_memory_buffer so that it returns a dds_loaned_sample_t and not just a raw loan pointer (that can't be used with dds_write because the buffer is not registered in the loan administration). With this change, the loaned sample (with custom size) can be used with the regular write function; only difference with the regular request_loan is that the user can specify the size in this case. The request_raw_loan fn prototype is removed from the PSMX endpoint ops; request_loan with the user-specified size is used instead. Signed-off-by: Dennis Potman <[email protected]>

…etting is used Signed-off-by: Dennis Potman <[email protected]>

…tion Signed-off-by: Dennis Potman <[email protected]>

... by adding SIGINT handler that terminates the process cleanly. Signed-off-by: Erik Boasson <[email protected]>

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Implementation of Publish Subscribe Message Exchange interface #1676

Implementation of Publish Subscribe Message Exchange interface #1676

Commits on Sep 19, 2023